Dual-Branch Spatial-Temporal Fusion Network for Dynamic UWB Localization in Longwall Mining

When measuring the shearer position in the underground mining face by ultrawideband (UWB), the non-line-of-sight conditions, complex noise, and signal attenuation always make it difficult to effectively reduce measurement noise. Therefore, it is hard to establish an accuracy model to obtain the shearer position directly. This article presents a dual-branch spatial-temporal fusion network (DB-STFN) with an encoder-decoder architecture to address this issue. In the encoder, a UKF-based dynamic kinematic model between the shearer and UWB anchors improves signal-to-noise ratio. A multiscale Transformer combined with multilayer perceptrons captures the spatial-temporal dependencies in the data, mitigating signal discontinuities caused by packet loss. The dual-branch decoder simultaneously estimates the positions of the shearer and hydraulic supports. The hydraulic supports branch maps spatial-temporal features to the global coordinate system for UWB anchor position accuracy, while the position of the shearer is computed from the flattened features. The intersection of UWB anchor data and labels representing the relationship between hydraulic supports and shearer is incorporated into the loss function to guide convergence. Experimental results show that DB-STFN achieves a shearer positioning RMSE of 0.1670m and a MAE of 0.1244m, representing a 42 % improvement over long short-term memory (LSTM) baselines. For anchor correction, the model reduces the mean error to 0.377m, ensuring spatial consistency across the dynamic environment. Robustness tests indicate that the system maintains a bounded error of 1.10m under severe synthetic NLOS bursts with biases of 3-13m, outperforming conventional filtering methods. Furthermore, the architecture contains 1.21M parameters and has an inference latency of 0.95ms on a standard CPU, supporting real-time deployment on resource-constrained industrial edge platforms. Ablation studies show that the joint optimization strategy reduces the positioning RMSE by 46.3 % compared to single-task configurations, while the shared temporal multilayer perceptron (MLP) and Huber loss enhance the system's resistance to measurement outliers.